SmartSustain Recommender System

• SmartSustain Recommender is a sustainable recommendation system that integrates environmental metrics and multi-objective optimization to deliver eco-friendly suggestions.
• The system employs diverse architectures, including edge-based, multi-agent, and retrieval-augmented frameworks, balancing user relevance with environmental impact.
• Evaluation focuses on metrics like carbon footprint, energy savings, and composite sustainability scores, providing transparent and actionable insights.

The SmartSustain Recommender designates a class of recommender systems that optimize for sustainability objectives—such as minimizing environmental impact while maintaining user relevance—across diverse domains. Characterized by integration of environmental metrics, user-centric design, and transparent algorithmic choices, SmartSustain Recommender systems explicitly address the need for sustainable consumer behavior, organizational decision support, and environmentally informed personalization. The term encompasses both architectural strategies (e.g., data-driven, edge-based, or multi-objective optimization) and the operationalization of evaluation metrics aligned with United Nations Sustainable Development Goals (SDGs) (Zhou et al., 12 Nov 2024, Felfernig et al., 4 Dec 2024, Said, 10 Jan 2025, Banerjee et al., 20 Oct 2025).

1. System Architectures and Algorithmic Design

SmartSustain Recommender systems exhibit considerable architectural diversity, employing different algorithmic frameworks tailored to their specific sustainability context. Typical approaches include:

• Two-sided platforms for food sourcing: Systems such as the Sustainable Recipes recommender connect food ingredient requirements to local organic producers by integrating recipe datasets with geocoded provider information. The architecture is modular, enabling input of a GPS coordinate and outputting a selection or recommendation of recipes and suppliers that collectively minimize food miles. Algorithms perform ingredient-producer matching via string matching and calculate food miles via Euclidean distance optimization (Herrera, 2020).
• Multi-objective optimization for value-sensitive baskets: Retail recommenders formalize the task as a multi-objective problem, minimizing for personal goals (taste similarity, nutritional needs, cost) and sustainability (GHG emissions, water footprint). Solutions are generated through evolutionary algorithms (e.g., NSGA-II, G3A) that support non-dominated recommendations aligning with both consumer values and environmental targets (Asikis, 2021).
• Edge-based and multi-agent architectures for energy management: In residential smart grid and IoT energy use settings, recommender systems utilize local data streams (from sensors and smart plugs) and process recommendations on resource-constrained, privacy-respecting edge devices. Algorithms extract “micro-moment” features from high-frequency sensor data and generate explainable energy-saving recommendations by associating user behavior patterns with optimal device scheduling (Sayed et al., 2021, Riabchuk et al., 2022).
• Collaborative and clustering methods for energy communities: Systems for renewable energetic communities cluster users based on historical consumption patterns (K-means on time-series energy use) to maximize within-community energy sharing, thereby reducing societal reliance on non-renewable sources (Guzzi et al., 2022).
• Retrieval-augmented generation frameworks for sustainable tourism: RAG-based pipelining with sustainability-augmented reranking (using city popularity and seasonality metrics) supplies personalized and environmentally informed city recommendations. LLMs are explicitly conditioned on sustainability during prompt assembly (Banerjee et al., 26 Sep 2024, Banerjee et al., 20 Oct 2025).
• Model-agnostic reranking for greener item recommendation: In e-commerce, post-hoc reranking adjusts the utility function of existing recommendation lists by linearly combining predicted rating and a normalized greenness (e.g., carbon footprint) score. This approach is modular and does not require modification or retraining of the base model (Kalisvaart et al., 21 Mar 2025).

2. Sustainability Metrics and Multi-Objective Trade-offs

Robust quantification of sustainability outcomes is central to SmartSustain Recommender systems:

• Energy and Carbon Footprint: Metrics include energy used in training, inference, and recommendation (ECRec, ECTrain), with CO₂-equivalence often derived from kWh using standardized conversion factors (e.g., 481 gCO₂e/kWh as global mean). Average carbon footprint of recommended items (AvgCarFI) and green item recommendation rate (GIRec) are used to assess environmental burden of output (Arabzadeh et al., 12 Oct 2024, Felfernig et al., 30 Jul 2025, Wegmeth et al., 16 Sep 2025).
• Composite Sustainability Scores: Some systems compute composite scores as weighted sums of transport efficiency, popularity, seasonality, and interest match (e.g., S = w₁·m₁ + w₂·m₂ + ...), with weights reflecting domain-specific priorities or personalization input (Banerjee et al., 20 Oct 2025).
• Greenness and Multi-metric Optimization: Utility and loss functions may integrate both relevance and sustainability. For instance, in a reranking context, the hybrid utility is H₍ᵤ,ᵢ₎ = α·f₍ᵤ,ᵢ₎ + (1−α)·gᵢ, where f is predicted relevance and g is the item's greenness (Kalisvaart et al., 21 Mar 2025). Multi-objective formulations are typical: $\min_{x}(J_1(x,x^*),...,J_M(x,x^*))$, with each $J_j$ reflecting a sustainability or personal goal (Asikis, 2021).
• Social and Economic Metrics: Additional axes include fairness (e.g., demographic parity in exposure), diversity (e.g., intra-list diversity), serendipity, and local business promotion rates. Cross-cutting measures such as long-term customer satisfaction and reduction in harmful exposures also feature in advanced evaluation frameworks (Felfernig et al., 30 Jul 2025, Felfernig et al., 4 Dec 2024).

3. Data Integration and Knowledge Modeling

A distinguishing property of SmartSustain Recommender systems is the integration of heterogeneous datasets reflecting both traditional utility and sustainability attributes:

• Food Sourcing: Data fusion involves recipe databases, geocoded supplier listings from sources like the USDA Organic Integrity Database, and supermarket locations (Herrera, 2020).
• E-Commerce and Consumption: Purpose-built datasets, such as RecipeEmission, provide CO₂-eq labels for each item, requiring extensive domain-specific scraping, ingredient standardization, and manual reconciliation of footprint data (Kalisvaart et al., 21 Mar 2025).
• Tourism and Urban Mobility: Datasets aggregate transport emissions, city popularity indices (number of POIs, visitor ratings), and seasonality derived from historical footfall to supply the feature set for recommendation and reranking algorithms (Banerjee et al., 26 Sep 2024, Banerjee et al., 20 Oct 2025).
• Energy and Smart Grids: Systems employ high-frequency, appliance-level time-series data, ambient context measurements, and real-time electricity pricing streams. Multi-agent knowledge bases and context-driven inference rules orchestrate device-specific recommendations (Sayed et al., 2021, Riabchuk et al., 2022).
• Sustainability Regulation and Reports: NLP systems operate on segmented corporate reports, using pre-trained BERT models for segment–regulation relevance assignment, often with regulatory standards (e.g., GRI) as the annotation schema (Hillebrand et al., 2023).

4. User-Centric Visualization, Nudging, and Explainability

Interface and feedback mechanisms are prioritized to facilitate informed, sustainable choices without undermining usability:

• Interactive Dashboards and Visualizations: City cards, animated progress bars (e.g., for CO₂e), real-time feedback banners, and radar comparison charts enable users to directly compare environmental trade-offs (Banerjee et al., 20 Oct 2025).
• Personalization and Adaptive Scoring: Systems capture fine-grained or survey-based personalization to dynamically reweight scoring metrics (e.g., elevating the weight of interest-match in the absence of personalization input) (Banerjee et al., 20 Oct 2025).
• Nudging Mechanisms: Context-aware banners and dynamic recommendations surface alternatives aligned with sustainability when users select high-impact options. Nudges are designed to gently steer, rather than enforce, behavioral change.
• Explainability and Transparency: Models provide granular explanations linking recommendations to sustainability metrics (e.g., “this route reduces CO₂e by X%,” or “local sourcing of basil minimizes food miles”), increasing user trust and supporting the system’s advocacy role (Hillebrand et al., 2023, Felfernig et al., 4 Dec 2024, Felfernig et al., 30 Jul 2025).

5. Empirical Effectiveness and Case Studies

Measurement of effectiveness involves both algorithmic performance and user paper analyses:

• Algorithmic Validation: Simulation studies and offline experiments consistently demonstrate that multi-objective and hybrid systems can achieve sustainability improvements with modest compromise in traditional metrics (e.g., SVD reranking with α=0.6 yielding 72.6% GNDCG@20 improvement for a 2.7% NDCG@20 loss) (Kalisvaart et al., 21 Mar 2025). Counterfactual simulations show that basket replacement strategies can reduce environmental impact across thousands of consumer trajectories (Asikis, 2021).
• User Studies: Preliminary interactive studies (n=21) for SmartSustain city travel recommenders indicate high usability and perceived sustainability impact by participants, with strong endorsement of visualizations and feedback features (Banerjee et al., 20 Oct 2025).
• Real-World Impact: Edge-based energy-saving recommenders deployed as part of Home-Assistant reduce both energy waste and privacy exposure. Industrial computation allocation frameworks (e.g., GreenFlow) have achieved up to 41% computation reduction, equivalent to ~5000 kWh and 3 tons of CO₂e daily savings, with little or no revenue loss (Lu et al., 2023).

6. Challenges, Evaluation, and Future Directions

Despite promising advances, SmartSustain Recommender systems face significant challenges:

• Data Scarcity and Quality: Scarcity of public datasets explicitly labeled for sustainability indicators (e.g., CO₂e, water use), as well as heterogeneity in source data, poses a barrier to generalization and benchmarking (Zhou et al., 12 Nov 2024, Kalisvaart et al., 21 Mar 2025).
• Evaluation Frameworks: New multi-objective architectures require metrics and protocols that quantify sustainability impacts alongside accuracy, engagement, and satisfaction. Standardized reporting of energy use, CO₂e (e.g., per kWh), and sustainability gains is advocated (Felfernig et al., 30 Jul 2025, Wegmeth et al., 16 Sep 2025).
• Environmental Cost of Modeling: Deep learning models, while often delivering state-of-the-art accuracy, exhibit energy and carbon footprints orders of magnitude above traditional methods, sometimes by factors up to 42x for equivalent pipelines (Wegmeth et al., 16 Sep 2025, Said, 10 Jan 2025).
• Trade-off and Personalization Mechanisms: Dynamic adjustment of consumer and sustainability weights, adaptive reranking, and personalization of nudges remain open areas for optimization—particularly for systems seeking wide adoption without user alienation (Kalisvaart et al., 21 Mar 2025, Jing et al., 19 Aug 2025).
• Broader SDG Alignment: Integration with SDGs, such as fairness (SDG 10), responsible consumption (SDG 12), climate action (SDG 13), and institutional trust (SDG 16), requires context-aware fairness, transparency, and accountability protocols embedded in system design (Said, 10 Jan 2025, Felfernig et al., 4 Dec 2024).

7. Prospects and Generalization

The SmartSustain Recommender platform is a paradigm for aligning recommender system research and deployment with environmental stewardship. Future systems are anticipated to deepen integration of sustainability attributes across domains (food, travel, energy), advance explainable AI that contextualizes environmental trade-offs, and systematize measurement/reporting of both direct model footprints and the indirect consequences of recommendations. Standardizing sustainability metrics and developing optimization frameworks that balance accuracy and sustainability will be central to this evolution (Felfernig et al., 30 Jul 2025, Zhou et al., 12 Nov 2024, Banerjee et al., 20 Oct 2025).